An empirical study to test the significant effect of bellwethers on predictive modeling in smart homes

Smart homes are envisioned and expected to be furnished with effective and reliable sensors for constant monitoring, sensing, and actuating the environment. Data from these sensors can be analyzed and used for the provision of various services by automating common tasks undertaken in the homes with minimal disruption to human activities. Thus, in order to provide these services, there is the need to obtain an intelligent system to accurately predict the most likely event to occur in a smart home. These intelligent systems rely on prediction algorithms trained from existing data for their functioning. The study seeks to invesigate the feasibility and significance of using Bellwethers for predictive modelling in smart homes. Recent works in the domain of software engineering – software defect and effort estimation have shown that the use of exemplary data (referred to as Bellwether) in the training set relatively improves prediction accuracy. These results were obtained based on an empirical analysis of the Bellwether effect i.e., the existence of Bellwethers in a given dataset. Existing studies in the context of smart homes focus on using all data from various sensors for training and validating the prediction algorithms for intelligent systems in smart homes. This mode of training and validating the prediction algorithms does not always yield accurate prediction results due to the variations and heterogeneous nature of the population dataset. A novel framework based on Markov chains and Deep learning algorithms is constructed to sample Bellwethers for predictive modelling in smart homes. It was observed that the sampled Bellwethers yielded improved prediction accuracy for the studied dataset as compared to using all data instances or population set (growing portfolio). It was evident that Bellwethers can improve the reliability and accuracy of intelligent systems built for smart homes.